Editing Learning (section)

===Ideas===
* Would it be possible to train a network on cut up sections based on how much they cause the neurons to spike? Don't train sections that have a big response already.
* Would it be possible to modify the inputs to 'censor' bits that cause over-fitting.
* Would it be possible to 'move' and object across the view overtime and learn that it's still the same object as a kind of data-augmentation?

* Could an auto-encoder be used to synthesise convolution filters for a pre-initialisation?

* Is it possible to learn a fitness function by starting in the goal state and trying to learn how to leave it.
* Then learn how to leave that state for a new one.
* Would be easier in a discrete, reversible, deterministic world.
** Would need to define how different 2 positions need to be to be considered different states.
** Is is possible to learn what a 'state' is?
** by taking 2 'non-goal' states and learning to move between them without triggering the goal and using a state halfway between as a new state?
** By taking a bunch of 'non-goal' states and and finding the maximum difference between them?
** Or using the distance between goal and non-goal?
** Or by using the distance between 2 non-goal states?
* Need to be able to reverse the 'move out of goal'.
** If the actions are reversible and deterministic then just undo them.
** Could relearn how to get from the state to the goal
** How to determine how far away a non-goal state is? How much time/how many actions it takes to get to the goal state from the non-goal state?
* What about a key/lock/door puzzle
** By default it wouldn't learn to put the key in the lock as the puzzle would either start in a solved state (or actor would be stuck behind the door).
** Could start the puzzle solved and make it become unsolved as the actor walks backwards, ie you pick up the key when you go though the unlocked door which becomes locked, then have to 'loose' the key it in the place where it's actually obtained. Is this just turning into as complex a problem as solving the puzzle in the first place?